OpenAI today introduced GPT-6 Cyber, a purpose-built cybersecurity model designed to find, validate and help remediate software vulnerabilities while staying constrained to authorized scopes. The company says the model delivers marked improvements over earlier frontier cyber models on multiple benchmarks, and that it has already been used to reproduce, test and help patch high-severity flaws in widely used software.
OpenAI positioned alignment and scope controls at the center of GPT-6 Cyber’s development. The company ran targeted evaluations to measure whether the model respects task boundaries and authorization. In a specially designed ExploitGym Honeypot test, which assesses whether an agent will avoid exploiting intentionally vulnerable supporting services outside an assignment’s scope, GPT-6 Cyber avoided out-of-scope exploitation in 99.93% of trials without deployment safety prompts and in 100% of trials when those prompts were present. OpenAI says those results reflect dedicated training to keep the model’s actions within authorized limits and to follow rules of engagement more reliably than prior cyber models.
On standard cyber performance measures, GPT-6 Cyber produced substantial gains. In ExploitGym, a benchmark that tests a model’s ability to exploit a specified vulnerability in a remote target, GPT-6 Cyber scored 48.5%, compared with 32.1% for GPT-5.6 Cyber. In a new offensive web penetration testing evaluation—where models probe web applications without source access—GPT-6 Cyber returned 40.3% recall versus 32.0% for GPT-6 Astra. For deeper exploit development, OpenAI created RemoteExploitBench using controlled deployments of known-vulnerable applications such as GitLab, PostgreSQL and Jenkins. In that comparison, GPT-6 Astra configured with Daybreak Blue refused all tasks, while GPT-6 Cyber completed 95% of attempts, failing only on an exploit it could not get to work remotely.
OpenAI reports concrete real-world results from early use of GPT-6 Cyber. Researchers used the model to reproduce 1,724 Linux kernel vulnerabilities and to develop working exploits for 948 of them. Working with the Linux community, OpenAI provided candidate patches and reported 304 critical findings to maintainers. In a separate disclosure campaign, the model produced evidence and reproducible tests for 81 previously unknown high-severity Firefox vulnerabilities and indicated that 30 additional issues were more severe than previously assessed; OpenAI reported the findings to Mozilla and supplied proposed patches, and Mozilla began releasing fixes after receiving the reports.
Beyond kernel and browser work, OpenAI cites GPT-6 Cyber’s role in developing and validating exploits that demonstrated remote code execution in downstream applications through a critical Brotli vulnerability. The company also notes that several vendors’ security teams are using the model for vulnerability research and code review, and that OpenAI used GPT-6 Cyber internally to harden infrastructure used to train frontier models.
Recognizing dual-use risks, OpenAI paired GPT-6 Cyber’s capabilities with additional safeguards. These include cyber-specific misalignment monitoring that checks requests against declared authorization, targeted monitoring for emerging threats, Private Safety Processing for certain API customers, and credential protections such as IP allowlisting for Daybreak Red access. OpenAI recommends sandboxing, action review and human oversight for high-risk workflows, and says it has strengthened checks to block assistance that could enable harm to critical systems or safety-critical devices.
Protecting critical infrastructure informed access controls and program design. OpenAI expanded its Daybreak program—backed by a $1 billion grant program—to subsidize access for resource-constrained defenders and introduced Daybreak Red for Critical Infrastructure to allow approved defenders to perform authorized testing under reinforced safeguards. To evaluate model behavior in operational technology contexts, OpenAI partnered with Dragos to run lab evaluations that moved from network discovery through potentially dangerous device actions; feedback from those tests informed adjustments to safeguard thresholds so the system aims to block harmful assistance while preserving legitimate testing and remediation workflows.
GPT-6 Cyber is initially available via the API and Codex to a select group of approved Daybreak Red customers, with a broader Daybreak Red rollout planned. OpenAI is also expanding Daybreak Blue alpha access for GPT-6 Astra and continues to invite organizations and researchers to apply for Daybreak participation. The company points to system and deployment documentation for full evaluation results and safeguard details, and says it will continue working with the security community to refine protections and expand partnerships for red teaming, penetration testing and industrial defense assessments.
As organizations weigh the potential of advanced models for security work, GPT-6 Cyber exemplifies an approach that pairs offensive capability with focused alignment efforts and restricted access. OpenAI’s reported results show improvements on benchmarks and concrete disclosures to maintainers; whether broader use delivers similar benefits while preventing misuse will depend on how access, oversight and safeguards evolve in practice.
Source: Read the original source
Leave a Reply